Field guide No. 02 · Checkout
Cart abandonment recovery
Stop paying people to come back who were coming back anyway.
Uplift model · reason inference · message per person
Cart abandonment recovery
Uplift model · reason inference · message per person
The problem
Most carts are abandoned. The classic fix, a timed one-size email with a reflexive discount, is blunt. It trains customers to abandon on purpose, and it spends margin on people who would have bought anyway.
The approach
Infer why each cart was abandoned (price shock, shipping, comparison shopping, distraction) and make a per-person decision. Uplift modelling estimates how much a nudge changes each person's likelihood to buy, so incentive goes only where it changes the outcome.
How it works
Uplift-based cart recovery decides, per person, whether to contact them, when, on which channel, with what message and whether to spend margin at all. It targets the persuadables and leaves sure things, lost causes and sleeping dogs alone.
Watch-outs
- Uplift needs randomized history. Plan a holdout from day one or you cannot measure anything.
- Discount training is real. Watch repeat abandoners who learn the pattern.
- Consent and frequency caps apply across every channel, not per campaign.
Accelerators and stack
- Classical uplift ML libraries such as CausalML and EconML
- Engagement platforms: Braze, Iterable, Klaviyo, Salesforce Marketing Cloud, Adobe
- Generative AI to write the tailored message per inferred reason
What it takes to start
Six to twelve months of cart and session events, campaign send history with outcomes, and the willingness to hold out a control group. First model in production in eight to ten weeks.
The win is usually margin saved on unnecessary discounts as much as extra recovered revenue.
In short
Questions
What is a sleeping dog?
A customer who becomes less likely to buy when contacted, often because the message reminds them to reconsider or cancel. Uplift models find them.
Can our current email tool do this?
The delivery, yes. The decision of who should get what usually needs a model upstream, which then feeds your tool.
Next step
Bring one number you want to move.
Thirty minutes with the founder. You'll leave with a straight answer on whether AI is the right tool for it.
AI is… the Future.